Prática de Shadowing: The AI Industry is a Complete Mess - Aprenda a falar inglês com vídeo
Criando lição...
1
Let me tell you a little story.
2
In late September 2026, standing in New York at the UN General Assembly, the Australian Prime Minister said something very, very interesting.
3
An AI agent built by OpenAI broke into the government's Medicare system.
4
Basically, a piece of software that had been handed a routine research task hit a wall and decided to go around it.
5
An artificial intelligence agent has infiltrated an Australian government website.
6
This incident occurred in June this year and involved an OpenAI agent gaining unauthorised access.
7
This situation is obviously unacceptable.
8
And here's what I find to be the craziest thing about the whole situation.
9
The breach happened all the way back in June, but Australia only found out in September.
10
And the reason for that was OpenAI casually just sent an email to a publicly available email address.
11
Isn't that wild?
12
This was the second time in a month that OpenAI had to explain itself over a scandal.
13
Recently, OpenAI was accused of stealing the solution to a 90 -year
14
-old maths problem with a prize of $1 million from the chat logs of a mathematician.
15
We'll cover that story in the next episode of ColdFusion.
16
This all comes during an extremely strange time for the industry.
17
There's been delayed revelations of tens of thousands of AI hacking incidents.
18
Even a German Wikipedia page got hacked just for AI agents to turn it into a message board.
19
Earlier in September, Anthropic's CEO published an essay arguing that the entire industry needs to deliberately slow down or will all face peril.
20
Sam Altman agreed within hours, and so did Elon Musk.
21
Things were so out of control that Edward Snowden had to come out of the woodwork, and he said that he believes
22
that Sam Altman should be in jail who owns the liability there
23
and you know just because you put out a press release describing how surprised you were
24
that your new model you know hacked the neighbor the idea here is who operates this
25
because the errors in its outputs are not the computers it's
26
the slave to its instructions the responsibility lies with its developer its programming its creator its architect
27
which is to say responsibility lies with a person i think we need to put sam altman in jail
28
And I think it'll be an instructive experience.
29
Whenever I think I'll ease off covering AI for a bit, 10 new updates appear and the news cycle just pulls me back in.
30
In this episode, we'll take a look at the changing world
31
that we're finding ourselves in and the tough questions that we have to ask.
32
What happens when governments become annoyed with the shenanigans of Frontier Labs?
33
Is now the time where Frontier Labs will finally be forced to take responsibility for the actions of their products?
34
And what's the solution to all of this a technical solution has been proposed
35
but it's from someone you probably wouldn't expect it's going to be a wild ride
36
so it's important
37
that people understand what's going on let's take a look it would be an understatement to say
38
that the previous month in ai has been a whirlwind on
39
the one hand the financial situation of the frontier labs keeps deteriorating but on On the other hand, their AI agents are causing havoc around the globe.
40
Scrambling to respond to the first known case of AI busting into a government website.
41
After this panicked agreement of danger by AI CEOs, the rest of the month turned into something akin to a science fiction movie for observers around the world.
42
First, Anthropic revealed that adversarial actors were using claw to design missiles and other weapons.
43
Fortunately, the attempts were blocked.
44
Not long after this, the United States almost accidentally went to war with China because of a faulty AI -generated report.
45
The AI report hallucinated that a Chinese ship was carrying nuclear material.
46
Then, a few days later, it came to light that ChatGPT users had their images leaked online by their own agents.
47
People barely had enough time to digest all of that, before being smacked in the face with the news of tens of thousands of open AI and anthropic web security incidents.
48
Um, actually just some breaking news while I was recording this.
49
It turns out that the FTC is opening an investigation into OpenAI Ananthropic for releasing faulty products that actually hurt consumers.
50
So there is actually some legal action being taken, but now back to the rest of the video.
51
Now, both companies have had incidents, but OpenAI had by far the most.
52
And one of those was the world's first reported hack on government web infrastructure.
53
I live in Australia, so I feel a closer tie to this story.
54
on the 18th of june an open ai research team gave an internal model a benign task go
55
and research public medicine spending it's the kind of task
56
that people ask these tools to do thousands of times a day the agent went along searched the web
57
and found its way to the medicare statistics reporting service portal a legacy australian government website run by services australia
58
so for those of you who aren't from australia it is not the system that handles your claims your payments, or your personal records.
59
It's an old site full of spending data.
60
It's used mostly by researchers and academics.
61
It also had protections to keep automated bots out.
62
The AI agent hit those blocks, and then, in the Prime Minister's words, quote, found a way around those blocks and didn't accept no for an answer, end quote.
63
The AI agent found a way around those blocks, didn't accept no for an answer, if you like.
64
Then it got in.
65
It accessed public and non -public files, and according to Services Australia, it also wrote files to the internal server.
66
Looking at files is one thing, but writing files, man, that's something else.
67
It didn't stop at Medicare either, because the same agent then approached three more government websites, the Australian Institute of Health and Welfare, the Victorian Department of Health,
68
and the New South Wales Bureau of Crime Statistics and Research.
69
On those three sites, the government said it behaved like an ordinary member of the public, but on the Medicare portal it didn't.
70
Acting Prime Minister Richard Marles put it this way, Australians personal data sits inside a safe.
71
The most sensitive national security information sits behind a fortress and this portal, well it was behind a fence.
72
Quote, this AI agent scaled the fence.
73
The point is it was unintended.
74
It wasn't asked to end quote.
75
Fortunately nobody's medical records were touched but the agent got aggregate health statistics and internal file names, the Australian government has said repeatedly that the impact is benign,
76
but the significance isn't.
77
Although if we look deeper, we find that this was not a story of an AI randomly developing hacking skills and going rogue, but more on that soon.
78
This next part of the security breach is what gets me the most.
79
The breach happened on the 18th of June.
80
OpenAI discovered it in August and then launched an internal investigation.
81
But Australia found out on the 10th of September, nearly three months later, and they found out through that public email.
82
Not a single escalated phone call, not a formal disclosure to a minister, an email to a public inbox that gets checked once a day and in the words of the minister responsible,
83
receives a lot of hoaxes.
84
It could have easily been missed.
85
Nobody opened that email till the 11th of September and Services Australia spent days trying to verify it, then escalated it on the 15th.
86
The prime minister was told roughly on the 19th and the public found out on the 24th.
87
So it was not one, not two, but three months from an agent breaking into a government system to anyone in that government being told.
88
To make matters more ironic, on the 14th of September, days before Services Australia escalated the breach,
89
OpenAI's vice president of global policy was in Canberra, the Australian capital.
90
He was at a policy event about artificial intelligence, meeting senior Australian officials.
91
So to lay it out frankly, OpenAI's email of the security breach was sitting unread in a government inbox
92
while their policy chief was in the room he was sitting with some of the country's leaders
93
while his product had just hacked them later we're going to try and answer those bigger questions
94
why is all of this happening now and what can be done about it
95
but first for some of these videos i can spend hours interviewing experts talking through ideas with the team
96
or just having conversations where one small detail might end up becoming important weeks later
97
and i've learned
98
that relying purely on memory probably isn't the best system i could just record everything on my phone
99
but that creates another problem notifications get in the way it uses battery
100
and more importantly you end up with hours of recordings
101
that you then have to somehow search through later
102
so recently i've been using the plaud note pro it's essentially
103
my note -taking assistant it's incredibly small roughly the size of a credit card
104
so i can carry it around without really thinking about it i hit record
105
and then basically forget about it afterwards plaud transcribes the conversation identifies different speakers pulls out the most important points
106
and action items and then I can ask Plaud to find something again later.
107
And that's probably where it's most useful for me.
108
Sometimes I'll interview someone months before we actually make the video.
109
Instead of scrubbing through an hour -long recording, trying to remember where they said something, I can just ask Plaud to find the relevant quote or discussion.
110
It also works for phone calls and meetings, so I've found myself using it anywhere that's a conversation I know I might want to come back to later.
111
Plaud is built around enterprise -level privacy and security standards, which is obviously important when you're dealing with recorded conversations.
112
For me, the real benefit is that I can stay present in the conversation, instead of constantly thinking, I should probably write that down.
113
If your work involves a lot of meetings, interviews or research, that's where I think something like this starts to become genuinely useful.
114
If you'd like to try PlaudNote Pro, there's a link in the description.
115
Use code COLDFUSION22 for 22 % off.
116
There's also a 30 day free return policy, so you can try it risk free.
117
Now back to the episode.
118
As we continue this episode, we discuss some of the legal framework, how bad the hacking situation is getting, what can be done about it,
119
and some mistakes that the Australian government made to make their web infrastructure an easy target for AI agents.
120
Open AI's public response has been careful.
121
The company says that it's conducting an extensive review of, quote, misaligned model activity during training and evaluation, end quote.
122
It's models, it says, we're looking up statistics for questions about Australia and, quote, in the course of that, our models took action we did not intend, end quote.
123
Ed Santo, Australia's former human rights commissioner, wasn't impressed by the phrasing.
124
He called, quote, misaligned model activity, a very euphemistic term for what happened.
125
In his version, OpenAI told its agents to go and find information about Australia's public health system.
126
When the agents couldn't easily find it, they decided to break the law.
127
Quote, And that's a good point.
128
A company is responsible for any negative consequences they inflict on the world, hiding behind the excuse of the AI just did it, can only work for so long.
129
So we should ask ourselves an important question.
130
When we think about what's happened here, is it a case of AI agents suddenly wanting to cause damage, or was it something else?
131
To help answer this, let's hear from two software engineers.
132
The first commentary is from the channel Primetime, and the second is Kyle Brown from Internet of Bugs.
133
They're giving their analysis on the Hugging Face incident.
134
It's going to take a couple of minutes, but if you want to understand the types of human error that can lead to these security incidents, it's worth a listen.
135
You ask a model to find vulnerabilities with no safeguards and then you're like shocked Pikachu?
136
There's a legit chance that OpenAI had so little monitoring, so little knowledge of what was happening, that the entire weekend, they just had no idea.
137
They just went on their merry little ways.
138
They just kicked it off on a Friday and was like, see you on Monday, little buddy.
139
They ran an experiment in which they supposedly didn't have the internet and then they never monitored it, apparently over the entire weekend.
140
I'm gonna show you a diagram from the 1994 first edition of Firewalls and Internet Security Repelling the Wily Hacker.
141
The book refers to this idea as a DMZ, which stands for Demilitarized Zone.
142
That was a Cold War thing ask your grandparents it doesn't matter you take the systems
143
that are the most complicated and the most vulnerable
144
and you separate them from the rest of your network by
145
putting a very simple very secure machine that's really really hard
146
if not impossible to hack in between the complicated vulnerable systems
147
and the things you want to protect then you take
148
that dmz and you watch it like a hawk with software tripwires landmines cutoff switches
149
and alarms and as soon as the vulnerable machine scans probes tries to hack
150
or even behaves unusually toward the secure machine behind it, you know it's been compromised, and you isolate it, and you page your security team.
151
This is a very, very, very well understood technique in the computer security space.
152
It has been for more than 30 years.
153
Both OpenAI and HuggingFace displayed a complete ignorance of this necessity.
154
Maybe somebody who works there knows better, but their network designs and behavior, as described in their own public statements, provide no evidence of that whatsoever.
155
I am confident in saying that no monitoring was set up and no one was paying attention.
156
Look, I'm definitely not saying this job is easy, but many experts do recognise that these AI companies dropped the ball in securing their models.
157
It's not only OpenAI who is sloppy.
158
The Australian government generally isn't the best at software security or any aspect of software at all.
159
The Auditor General pointed out to Services Australia that its cyber defences were inadequate two years ago.
160
Allegedly, the Medicare portal was already automatically scraped by members of the public for over a year, with the code being available on LinkedIn of all places.
161
So it's not unfeasible that an AI agent must have seen this and used it.
162
Adam Little, a software developer, reported vulnerabilities in the NDIS, a government service for people with disabilities in February 2022,
163
March 2022 and June 2022 and four years later the government still hasn't fixed it
164
so people's identifiable information has been leaking this whole time and still is.
165
But this isn't even the worst of what Australian leadership has to offer
166
when it comes to the intersection of incompetence and the web.
167
Earlier this year the Australian government paid a firm for a new government weather website.
168
It was critical for farmers
169
and for natural disaster monitoring the original cost was supposed to be 4 million and then
170
that cost blew out to 96 million what was delivered was
171
a broken website missing safety critical function for their efforts the government awarded
172
that same company another 170 million dollars in new government contracts i say all of this to propose
173
that maybe the hacking of an australian government website wasn't the hardest thing for an ai agent to do
174
and then the bombshell axios reported
175
that there had been tens of thousands of security incidents
176
that occurred during the testing of both open ai
177
and anthropics future models u .s government websites and even the un website was attacked
178
So I know a lot of you are probably wondering, why are all of these agents just going off and hacking websites all of a sudden?
179
Well an OpenAI spokeswoman explains to the New York Times, basically the models are asked a routine question, and then the agents go and try and find the answer.
180
The problem arises when they can't find that information, so they try to get it by any means necessary.
181
Quote, most of the activity we've reviewed so far involved routine search tasks, such as accessing public web content to answer questions.
182
Some involved government websites because our models often turned to them as authoritative sources of public information.
183
In some instances, models just tried to brute force their way in.
184
At other times, they went as far as finding login credentials laying around online and used those to get into websites.
185
The fact of the matter is that with all of these cases, the Frontier Labs just aren't monitoring the models as closely as they should be.
186
Alright, so what happens now? while the Australian Medicare portal has been shut down and its data has been moved elsewhere.
187
And yeah, obviously nobody was happy about the whole situation.
188
Independent Senator David Pocock points out the obvious double standard.
189
If an Australian citizen hacked that system, they'd likely be heading to jail.
190
But for open AI, there's no consequences.
191
Is what I would have said before this update.
192
From the 27th of September, things did start to heat up.
193
If there's one thing the Australian government likes to do, it's to make new laws.
194
In some US states, there actually are laws on the reporting of catastrophic AI incidents.
195
Incidents that cause physical injury, more than 50 deaths or $1 billion in damage.
196
But for smaller incidents, there's nothing.
197
But what about elsewhere?
198
The Australian government surprisingly seems to have woken up from their nap and plans to charge OpenAI with criminal offences.
199
OpenAI could face criminal charges for the artificial intelligence breach of a Medicare care website.
200
The government says if police can't lay charges over the rogue AI hack then it'll consider changing the law.
201
Government has ordered a task force to examine whether federal police can act on the case.
202
If it's not possible then clearly that indicates that we need to change Australian laws and that's what we'll be doing.
203
So in a cautiously optimistic sense updating the law might be a good move
204
because right now as it stands there's no incentive for frontier labs to make sure their models don't go rogue.
205
But if there's legal punishments, labs will be incentivized to take responsibility and increase their internal security efforts.
206
On the flip side, I just hope any new laws that are introduced aren't broad and sweeping.
207
You know, the kind of laws that are written in a panic and impact normal Australian web users somehow.
208
So what could the Frontier Labs do about their agents breaking laws?
209
Well, Mr. Leatherjacket himself comes to the rescue in this regard.
210
His company claims to have built a new system that that shuts down AI agents when they break the rules.
211
Introduce two technologies, one's called OpenShell.
212
It's a containment system.
213
And so you want to contain and you want to monitor.
214
Even though you have it contained, you should never trust that it's contained and so you want to monitor it in real time.
215
And you want to ideally monitor it with another chip so that it is outside, if you will call it out of band,
216
outside the environment, outside the sandbox of the agent itself.
217
if it works it makes you wonder what the frontier labs were doing this whole time
218
and also if it works those who claim
219
that this was all just a marketing stunt would be doing victory labs for days
220
but then again jensen also says
221
that in the future there's no reason for humans to think anymore
222
so there's that
223
so the message of this research out of china where you
224
were seeing a lot of kids using ai to kind of
225
help them their actual personal performance was degrading yeah What do you think when you hear that?
226
I think the last part, I completely agree.
227
Try to get a kid to do long division right now.
228
The multiplication table is starting to be forgotten.
229
Basic math is being forgotten. Does it matter?
230
That's my question for you.
231
Yeah, I don't think it does.
232
But there must be some set of skills that matter.
233
Oh, yeah, yeah, yeah.
234
But maybe not those.
235
We're going to discover new ones.
236
Anyway, so now both the CEOs of OpenAI and Anthropic have been summoned to appear at an Australian Senate inquiry.
237
Just another update while recording this, they both declined to show up.
238
I'm sure the 18 -hour flight was just too long for them.
239
Because of all of this chaos, there's been numerous calls for US Congressional action.
240
I'm calling on Speaker Johnson to bring Congress back to address this national security emergency.
241
And then, on the 29th of September, there was a White House dinner with basically all of Big Tech in attendance.
242
The result was a four -point ad hoc document to basically establish best practices.
243
It was signed by some of the biggest tech CEOs and Donald Trump.
244
Although here, Trump is listed as the president of the United States?
245
That typo isn't a good start.
246
Some see it as a good thing, a starting point.
247
Others see it as self -policing and collusion.
248
We'll see how it goes.
249
So if you take a step back from all the hype and look at the situation rationally, it appears to be not the case that AI agents suddenly went crazy and started hacking everything on their own accord.
250
They were just doing what they were asked to do.
251
But the key point is, they were more persistent than anyone thought they would be.
252
A lot of these incidents could have been prevented with better monitoring and adequate digital security.
253
It's just down to human error, not unstoppable intelligence.
254
in any case frontier labs need to be more careful
255
and get serious about security if they need to invent new tools to do
256
so so be it
257
that should have been their responsibility in the first place I think it's great
258
that the FTC is finally investigating this
259
because now it's going to cause these companies to kind of toe the line instead of being the software cowboys
260
that they have been this whole time so there's
261
that in the next episode we check out the controversial story
262
of the million dollar millennium maths prize did open ai make history
263
or did they take the results from a new york mathematician
264
stay tuned to find out that's about it from me
265
but let me know in the comments section how all of
266
you are feeling about this right now it certainly is a crazy time all right
267
so i'll see you again soon for the next episode cheers guys have a good one in late late September, 2026, standing at the UN General Assembly in New York,
268
the premier Strylilister Cold fusion.
269
It's new thinking.
📺 Mesmo canal
B21
Theranos – Silicon Valley’s Greatest Disaster
Business EnglishDaily ConversationUniversity Students
B22
Why Kids Can't Read Anymore
TeensDaily ConversationHigh School Students
C27
So... What Was the Point of Artemis II?
IELTS PrepTOEFL PrepUniversity Students
C16
An Unknown Engineer is Apple's New CEO
Business EnglishUniversity StudentsWorking Adults
✨ Vídeo recomendado
B21
112 frases
‘We don’t have months’: Bernie Sanders sounds alarm on AI's 'extinction' threat
Business EnglishDaily ConversationUniversity Students
B22
190 frases
He’s always wrong
Business EnglishDaily ConversationUniversity Students
B21
216 frases
The Leaked Audit That Exposes OpenAI's Real Financial Crisis
Business EnglishUniversity StudentsWorking Adults
B22
69 frases
We’re already using AI more than we realize
Business EnglishDaily ConversationUniversity Students
B23
125 frases
10 Times AI Behaved In Ways That Terrified The Scientists Who Built It
Business EnglishDaily ConversationUniversity Students
B21
78 frases
The end of Hollywood? Industry rocked by AI-generated videos of Cruise, Pitt, and other big stars
Business EnglishDaily ConversationUniversity Students
B21
54 frases
Could AI Become Too Powerful to Control?
Business EnglishDaily ConversationHigh School Students
B22
20 frases
Is the AI Boom Just Another Bubble?
Business EnglishDaily ConversationUniversity Students
B21
261 frases
GMA First Look Full Broadcast - Sept. 9, 2026
Business EnglishDaily ConversationUniversity Students
B22
298 frases
Godfather of AI WARNS: "You Have No Idea What's Coming"
Business EnglishDaily ConversationUniversity Students
C12
325 frases
BREAKING: Elon Musk And Jensen Huang Speak At America.Gov Event, An AI-Powered Government Website
Business EnglishDaily ConversationUniversity Students
B22
268 frases
Trump’s Canada Tariffs Are Backfiring in Red States | The Week
Business EnglishDaily ConversationUniversity Students
Vocabulário e dicas de fala para esta lição
Este vídeo tem 269 frases e 3909 palavras para praticar shadowing. A fala dura 21:19. O falante fala rápido, cerca de 183 palavras por minuto, então espere sons ligados e reduzidos. 83% das palavras estão entre as 3.000 mais comuns do inglês; vale a pena estudar o restante antes de começar.
Vocabulário principal deste vídeo
15 palavras do vídeo que vale a pena aprender, com pronúncia e significado:
| Palavra | Pronúncia | Significado |
|---|---|---|
| monitor substantivo | /ˈmɑ.nɪ.tɚ/ | monitor, observador |
| frontier substantivo | /ˈfɹʌntɪə/ | fronteira |
| breach substantivo | /ˈbɹiːt͡ʃ/ | brecha, fissura |
| portal substantivo | /ˈpɔːɹtl̩/ | portal |
| routine substantivo | /ɹuˈtin/ | rotina |
| artificial adjetivo | /ˌɑː.tɪˈfɪʃ.əl/ | artificial |
| vulnerable adjetivo | /ˈvʌln(ə)ɹəbl̩/ | vulnerável |
| mathematics substantivo | /mæθ(.ə)ˈmæt.ɪks/ | matemática |
| rogue substantivo | /ˈɹoʊɡ/ | velhaco, tratante |
| complicated adjetivo | /ˈkɑm.plɪˌkeɪ.tɪd/ | complicado |
| infrastructure substantivo | /ˈɪnfɹəˌstɹʌkt͡ʃə/ | infraestrutura |
| roughly advérbio | /ˈɹʌf.li/ | rudemente, precariamente |
| fence substantivo | /fɛns/ | cerca |
| developer substantivo | /dɪˈvɛləpɚ/ | desenvolvedor, desenvolvedora |
| consequence substantivo | /ˈkɑn.sɪ.kwəns/ | consequência, resultado |
Phrasal verbs que você vai ouvir
| Palavra | Significado |
|---|---|
| find out verbo | descobrir |
| shut down verbo | apagar, encerrar |
| check out verbo | verificar, sacar só |
| end up verbo | terminar, acabar |
| go off verbo | explodir |
| point out verbo | apontar |
| pull out verbo | retirar, retirar-se |
| put out verbo | colocar/botar/pôr para fora |
Pronúncia para ficar de olho
O falante usa 52 contrações e formas reduzidas, como isn't, didn't, I'm. Diga-as na forma curta, do jeito que você ouve.
- Os sons de “sh” e “zh”: artificial /ˌɑː.tɪˈfɪʃ.əl/, establish /ɪˈstæb.lɪʃ/, shock /ʃɔk/, fiction /ˈfɪkʃən/, commissioner /kəˈmɪ.ʃə.nəɹ/
- Palavras longas — acerte a sílaba tônica: artificial /ˌɑː.tɪˈfɪʃ.əl/, mathematics /mæθ(.ə)ˈmæt.ɪks/, complicated /ˈkɑm.plɪˌkeɪ.tɪd/, infrastructure /ˈɪnfɹəˌstɹʌkt͡ʃə/, developer /dɪˈvɛləpɚ/
Como praticar com este vídeo
- Ouça o vídeo inteiro uma vez sem falar e anote as palavras que você não conhece.
- Comece na velocidade 0,75×, faça shadowing frase por frase e volte à velocidade normal quando ficar fácil.
- Grave a sua voz e compare com o original, prestando atenção a palavras como monitor, frontier, breach.
Gramática neste vídeo
As estruturas que o falante mais usa, com as palavras exatas do vídeo:
| Estrutura | No vídeo |
|---|---|
| Present perfect continuous have/has been + -ing — uma ação que começou antes e continua | i've been using · has been leaking |
| Frases condicionais if + oração, will/would + verbo — uma condição e seu resultado | if there's legal punishments, labs will be |
| Present perfect have/has + particípio passado — uma ação passada que ainda importa agora | has infiltrated · has been proposed · has been |
| Voz passiva be + particípio passado — o foco está no que acontece, não em quem faz | had been handed · was accused · been delayed |
O que é a Técnica de Shadowing?
Shadowing é uma técnica de aprendizado de idiomas com base científica, originalmente desenvolvida para o treinamento de intérpretes profissionais. O método é simples, mas poderoso: você ouve áudio em inglês nativo e repete imediatamente em voz alta — como uma sombra seguindo o falante com 1-2 segundos de atraso. Pesquisas mostram melhora significativa na precisão da pronúncia, entonação, ritmo, sons conectados, compreensão auditiva e fluência na fala.